PostgreSQL backend

MLflow’s backend store holds every run, parameter, metric, tag, and the model registry. This chart enables the bundled PostgreSQL by default because the alternative loses all of it on restart.

mlflow:
  postgresql:
    enabled: true
    image:
      tag: "17.5.0"
    auth:
      username: mlflow
      database: mlflow
      existingSecret: ""
    primary:
      persistence:
        enabled: true
        size: 8Gi

When postgresql.enabled is true the upstream chart wires PGHOST, PGPORT, PGDATABASE, PGUSER, and PGPASSWORD into the MLflow container and sets --backend-store-uri=postgresql:// on the server. Nothing to configure by hand.

The credentials secret

kubectl create secret generic mlflow-pack-postgresql \
  --namespace mlflow \
  --from-literal=password="$(openssl rand -base64 32)" \
  --from-literal=postgres-password="$(openssl rand -base64 32)"

mlflow:
  postgresql:
    auth:
      existingSecret: mlflow-pack-postgresql

The name and keys are exact

• Name — <release-name>-postgresql. A release called mlflow-pack needs mlflow-pack-postgresql.
• Keys — password (the mlflow user) and postgres-password (the superuser). Both are required.

A missing key is the legible failure: the PostgreSQL pod never starts, stuck in CreateContainerConfigError naming the key it could not find. A misnamed secret is worse. The MLflow container reads PGPASSWORD from a secret whose name the upstream chart hardcodes to <release>-postgresql, and marks it optional: true — so if your secret is called something else, PostgreSQL comes up happily against it, MLflow starts with no password at all, and the only symptom is an authentication failure when it connects.

Inline passwords

For a throwaway cluster:

helm install mlflow-pack . --set mlflow.postgresql.auth.password=dev-only

Never in production, and never in a values file committed to a GitOps repository — Helm values are readable by anyone with access to the release.

Rotating the password

Update the Secret, then restart both workloads:

kubectl -n mlflow rollout restart statefulset/mlflow-pack-postgresql
kubectl -n mlflow rollout restart deployment/mlflow-pack

Changing the Secret alone changes nothing — the values are read into the environment at pod start.

Storage

The default 8Gi PVC is generous for metadata: runs, params, and metrics are small rows. Growth comes from run volume, not model size — models live in the artifact store.

Set the storage class explicitly on any cluster where the default is not what you want. PVCs cannot be reassigned to a different class after creation:

mlflow:
  postgresql:
    primary:
      persistence:
        size: 20Gi
        storageClass: fast-ssd

Migrations

mlflow:
  backendStore:
    databaseMigration: true
    databaseConnectionCheck: true

databaseMigration runs MLflow’s schema migration at startup, so a chart upgrade that bumps MLflow also updates the database. databaseConnectionCheck adds an init container that waits for PostgreSQL to accept connections, avoiding a crash-loop while the database is still starting.

Migrations are one-way

Rolling the chart back to an older MLflow does not roll the schema back, and an older server may not start against a newer schema. Snapshot the PostgreSQL volume before a major version bump.

Backups

The chart does not back up the database. On a Nebari cluster with longhorn-backup-pack installed, the PostgreSQL PVC is covered by the cluster-wide schedule if it is on the default StorageClass — worth confirming rather than assuming:

kubectl -n longhorn-system get volumes.longhorn.io \
  -l recurring-job-group.longhorn.io/default=enabled

For a logical backup:

kubectl -n mlflow exec -i statefulset/mlflow-pack-postgresql -- \
  pg_dump -U mlflow mlflow > mlflow-backup.sql

No -t there: a TTY rewrites line endings in the redirected dump.

A complete restore needs both halves — the database and the artifact bucket. Restoring one without the other produces runs that reference missing models, or orphaned files no run points at.

Using an external PostgreSQL

Disable the bundled instance and supply a connection string:

mlflow:
  postgresql:
    enabled: false
  backendStore:
    postgres:
      enabled: true
      host: postgres.example.com
      port: 5432
      database: mlflow
      user: mlflow
      # password: ...  — or omit it and set
      # backendStore.existingDatabaseSecret.{name,usernameKey,passwordKey}

Check the community chart values for the exact shape — that block belongs to the subchart, not this one.

The SQLite fallback

mlflow:
  postgresql:
    enabled: false

MLflow then falls back to SQLite — and not to a file, either. The upstream chart’s backendStore.defaultSqlitePath is :memory:, so the server runs with --backend-store-uri=sqlite:///:memory:. Everything is lost when the process restarts, along with the artifacts, and with more than one server worker each would see its own empty database. Fine for a five-minute demo or CI — the chart’s own test workflow uses it — and nothing else.

examples/standalone-values.yaml uses this combination deliberately; see Standalone deployment.